AI support triage for the support inbox. Every message is classified on arrival; the AI acts alone only where it has earned the right to, and hands the decision to a person when it hasn't.

Website · What works today · Español

What it does

• Classifies every message on arrival — type, priority, urgency and tone — before anyone opens the inbox.
• Orders the queue by ticket type, customer plan, mood, waiting time and recurrence, so the first screen is the work order.
• Earns its autonomy per customer. A ticket type is auto-approved only after measured accuracy, and a correction takes that right away again.

How a message becomes a ticket

flowchart LR
  A[Incoming message] --> B{Filter rules}
  B -->|spam, automated, bulk| X[Skipped · reason stored]
  B -->|relevant| C[AI classification]
  C --> D{Autonomy earned<br/>for this type?}
  D -->|yes| E[Acts · type, priority, position]
  D -->|no| F[A person decides]
  F --> G[Correction recorded]
  G --> D

Loading

Status

State	What it covers
Working today	Gmail ingestion · classification · priority queue · replies in-thread · internal notes · escalations · response-time targets
Built, off by default	Knowledge proposals · ticket assistant · second model · contact drafts · acknowledgement email
Not built	Channels beyond Gmail · billing · auto-resolution

What works, what is switched off, and where the marketing site promises more than the product does: docs/what-is-kairo.md.

Quickstart

Requires Bun 1.3+, a Supabase project, Google OAuth credentials, and an AI Gateway key (or local Ollama).

bun install
cp .env.example .env.local   # fill in the values
bun run dev                  # api :3001 · dashboard :5173 · landing :3000 · kelan :3002
bun run dev:inngest          # Inngest dev server for the pipeline (separate terminal)

bun test        # unit and integration tests
bun run build   # turbo build across the monorepo

Every variable is validated at startup by @kairo/env; see .env.example for the full list.

Architecture

apps/
  api/          Bun + Hono · REST v1 + Inngest pipeline (tiers, polling, retries)
  dashboard/    Vite + React 19 · the cockpit
  landing/      Next.js 15 · marketing site, auth, onboarding wizard
  kelan/        Next.js 15 · internal admin
  mobile/       Expo · scaffold only
packages/       shared libraries; the AI layer is intelligence/ (AI SDK, model roles, prompts)
supabase/
  migrations/   schema, RLS, append-only history

Evals

The part worth reading: a bench of hand-labelled emails scored against the same pipeline production runs, per field rather than pass/fail. Several design decisions in this repo — trusting the mail server's spam verdict, splitting buyers from sellers, auto-approving only support at first — came out of these numbers.

bun run eval:pipeline   # emails through the classifier → CSV
bun run eval:metrics    # F1, calibration and tone inflation per label

The corpora are private and are not in this repository.

Docs

Document	What it covers
what-is-kairo.md	What the product does today, plainly
que-es-kairo.md	The same, in Spanish
ai-engineering/	Capability map, skills and tools
observability.md	Langfuse + ClickStack
versioning.md	How rules and prompts are versioned

Contributing

Open an issue before a large change. Commits use Conventional Commits; user-facing strings need both English and Spanish.

License

Apache-2.0.